Agreed. I do agree with your point that they are very replaceable, but the amount of people I know that still only have tried ChatGPT makes me think that most users are quite lazy once they have a tool. I suspect the (vast?) majority don't bother moving without good reason.
So to me it seems to make sense to take as much marketshare in this as possible, rather than saving a few $10/100ms on COGS optimisation, then do the COGS optimisation later.
FWIW all LLM stuff is quite easily replaceable, but that's not stopping anyone from trying to aggressively grow marketshare.
Fair point, I always forget how much of an early adopter I and the developers I am surrounded by are. There are plenty of developers who use whatever the CTO approved at their shop, which yesterday was probably github's copilot but tomorrow could be Cursor + Cursor Agents if they continue to out execute.
So to me it seems to make sense to take as much marketshare in this as possible, rather than saving a few $10/100ms on COGS optimisation, then do the COGS optimisation later.
FWIW all LLM stuff is quite easily replaceable, but that's not stopping anyone from trying to aggressively grow marketshare.